The questions on the final will be drawn from the entire semester. For examples of questions covering the material presented before midterms 1 and 2 check out these links:

• Sample Exam 1 Questions
• Sample Exam 2 Questions

There will also be questions related to the papers presented by the students in class. At most 80 points of the final will relate to these papers. Below are some potential questions based on the papers presented by students. The questions regarding those papers will be drawn from THIS list:

1. How does the Sequential Minimal Optimization (SMO) method for training a
   Support Vector Machine work?  Make sure to discuss how the notion of the
   support vectors figure into this, what is meant by the term "smallest
   possible optimization problems" (and what such a problem represents), how
   the method chooses from amongst these problems (and how that choice affects
   the solution method).  You should also compare the SMO method to the
   Chunking method for training SVMs.

2. How does the method for Support Vector Clustering introduced by Ben-Hur
   et al. work?  What does the support vector machine attempt to do in this
   method?  Also, how is the graph produced from the resulting machine and
   what part does the connected components algorithm play in producing the
   clusters?

3. Kondor and Jebara introduced a kernel function that would be useful in
   what types of situations?  Give an example of such a situation.  What are
   the advantages of this kernel over other kernels for comparing data?

4. Explain the basic idea behind Principal Component Analysis (PCA)?  In what
   types of domains would be PCA be very effective?  What type of problem
   did Suykens et al. propose to solve to perform PCA and what mechanism
   do they use in solving this problem?

5. Explain how Valentini and Dietterich proposed to control the Bias-Variance
   tradeoff in SVMs and how they proposed to use the resulting SVMs in an
   ensemble learning method.

6. In the Jaakola et al. paper "A discriminative framework for detecting
   remote protein homologies" explain what is meant by the term remote
   homology?  In this paper Jaakola et al. proposed to use compare generative
   models - what type of model and how would this models be compared (what
   mechanism would be used in the comparison and what would be compared)?

7. Vinokourov et al. proposed a method that would allow (amongst other things)
   for cross language queries (a query in one language that could find a 
   relevant document in another language).  Explain the basics of their
   method and how such a query would be possible.

8. In the Bayes meets Bellman paper, Engel et al. proposed to learn a policy
   function as a Gaussian process.  What functions did they propose to learn
   to produce a policy (and what assumptions are built into this approach)?  
   What are the advantages of this approach?  What do they mean by the term
   on-line sparsification and how is it important to their approach?

9. What is meant by the term "model" in Wang and Dietterich's "Model-based
   policy gradient reinforcement learning"?  How is this model learned and
   how is it used in their approach?  What are the advantages and disadvantages
   of this approach with respect to other reinforcement learning methods?

10. What is the difference between an Markov Decision Process (MDP) and
   a Semi Markov Decision Process (SMDP)?  Explain how this relates to the
   use of temporally extended actions such as those introduced by Options
   (Sutton, Precup and Singh)?

11. What is an Option as introduced by Sutton, Precup and Singh?  How are
   options used in a reinforcement learner?  How can an option help speedup
   learning?

12. What is a Hierarchical Abstract Machine (HAM) as introduced by Parr and
    Russell? Who creates HAMs?  How does a reinforcement learner work in
    a system involving HAMs?  What is a choice point?